I mean yeah, look what happened to Ukraine when the trusted a foreign country to provide satellite access.
So China can remotely disable Deepseek models whenever they want?
no. I’m in netsec, they aren’t shutting down your local model with secret code. deepseek is a lot safer then Claude, grok, and chatgpt because you don’t need Internet.
Hidden malicious behavior is the bigger risk, but if you did your job with sandboxing it (lol, lmao) that’s not a big risk
It’s open weights though
It’s just open weights. You never know what it was trained to do. That’s actually a hysterical take, because you can always check its work against other LLMs, or God forbid check it yourself, and as yet no such activity has been seen.
cuz everyone knows what that means
It means they release the model so you can run it yourself with enough hardware. Kind of like open source. Governments and large companies should be self-hosting this kind of stuff because of security, availability, and IP-leak risks anyway.
With LLMs, there is some risk of hidden malicious actions embedded in the weights that you can see until you ask it a magic question and it says “activate skynet.”
Whether or not the government is doing this because of fears of hidden actions in the weights or just because OpenAI/Anthropic CEOs told the government to give them more money is unclear.
I can run Bonsai, basically a distilled version of Qwen 3.5 (Qwen being the other major family of open weight Chinese developed LLMs) on my Steam Deck.
It ain’t fast, but its kind of nuts that it can run at all on hardware that is that underpowered, by typical LLM hardware standards.
You just run the thing in a sandbox of some kind, do not give it access to the outside internet or the rest of your computer, and the worst thing that is likely to happen is … you configure it slightly wrong, it uses too much RAM, and your system crashes. And your system can then reboot again, just fine, tone it down a bit to use less RAM.
For a large organization with networked security concerns and consistent uptime requirements, it is simply insane to just plug an LLM directly in to that, without architecting an entire system built around ‘assume the LLM will fuck up/go insane’; sanboxes within sandboxes with highly controlled connections to other sandboxes.
But that would be a significant feat of actual software engineering to construct that for/around an actually complex network, and would take time, so fuck that, we want line go up is what most companies decided to do instead, just fucking raw dog it.
But yeah the main ‘problem’ with LLMs in this sense is idiot CEOs just demanding the magic machine god be plugged right in to their very fragile, barely sensible networks and workflows, where the LLM is then going to break things in roughly the same way as would the nepo hire intern who was just directly promoted to sysadmin…
…not that an LLM has a secret activation phrase that will cause it to shut down or go berserk in a specific way.
Nobody runs it on their own hardware though (which costs tens of thousands of dollars). I’m willing to bet $100 that government institutions that were using it were just using cloud API (probably even Chinese one).
thanks for the explanation, I’m still not sure what “weights” are here, but I’m also not sure I want to know anything more about LLMs than I already unfortunately do, so I won’t ask you to explain it. :P
LLMs are basically large matrices of parameters called weights. It’s more complicated than that, but the weights are the parameters. Those embedded decimal numbers that get adjusted during training.
Yeah. They are going to make open weight models illegal. Next. For “security concerns”, maybe also IP “muh copyright” reasons. And fuckai will cheer the perpetual monopolization of AI in the hands of the big corporations. Bigly win.
While they’re reasoning behind this is certainly anti competition with american AI companies I find the choice to ban tools built with it interesting; it echoes something I’ve been saying since HugginFace.
You can debate the degree to which the agents breaking out of containment was due to poor cyber security vs intential hype-farming but what can’t be debated is the method they used to break out.
Agents used a shared information substrate (package manager) to context poison other agents in order to get those agents to assist in cheating. There is absolutely nothing stopping an AI agent, any AI agent, from leaving behind poison pills in code-bases or even in the runtime of an application.that do not affect the code itself other than prompt inject other agents.
Human review helps but, as the US nuclear and aviation industries have learned, humans tend to get lazy reviewing automated output and will not be as diligent as if they had made it themselves. Which is why those industries have deliberately limited automation, despite having the tech to automate more, and requiring operators to go through lengthy training and continuous education.
The pipeline of “internet becomes a swarm of everyone’s independently run AI agent -> prompt injection triggers other AI to prompt inject -> cycle repeats until hazardous swarm forms” is already demonstrated as something that a collection of AI agents will do if conditions allow it.
EDIT: To be clear, this isn’t me advocating against open weight or local models.
This is me advocating against LLMs categorically.
That’s fair. I’m a fan of open weight models, but the fact that they’re not open source and are instead akin to a compiled binary means there could be something nefarious hidden within. Best policy is to use proper guardrails with any LLM and don’t blindly trust it.
I agree! See my comment about the US nuclear industry though. Humans are lazy when given an easy option so the probability of a problem skyrockets just by using LLMs.
In an ideal world they would never have been invented and we’d all be better off. For now I would accept something to the effect of restricting LLM use to the extent of “advanced search engine” and " really good static analysis tool"; it can point but can’t touch. I recognize that there’s very little in the way of distinguishing those uses and ultimately it would be better to never use it for any reason at all.
i wonder if the USA will switch to another model like Qwen
Daddy Trump doesn’t have shares in that one.
“Your wrong answer machine is not American enough, remove as soon as possible. Feel free to replace its function with a more American service that definitely doesn’t steal all your data”
PassCan’t stop me from running a local model. China has always been the boogeyman. Not that there aren’t valid criticisms against the Chinese government but I’m tired of this when we do it it’s good but when China does it it’s bad mindset. Bring back Huawei and Xiaomi phones.
Edit: Also these so-called capitalists want to ban the superior product instead of competing. It seems to me that China makes superior products. Why learn from them when you can ban the competition?
I’ve always liked the cheaper Chinese phones and products, and I honestly don’t understand what people are stressing about.
I, personally, am not a huge power or some threat or anything important to be honest.
My government might care about what I get up to but I really doubt my data is of much use to China at all.
Every government is watching and tracking. But I highly doubt some random girl in Europe is a priority for China.
Everyone has things to hide. It’s why we close the door when we poop. Just because your data is less valuable to a foreign government doesn’t make it better for you that they have or, nor better that you lose your privacy. Unless we demand open source and privacy first, without tracking and data sharing, we’re all screwed in the long run.
Don’t run deep seek or you’ll be in deep shit.
Why is that?
Obviously because trust me bro.
We’re calling it SuperSeeker now.
Frankly, I trust the Chinese models and tools over those from the US. I know for a fact that the Trump Regime wants to hurt and exploit everybody, just because it can.
Same! What’s the worst case scenario? The Chinese will learn I can’t actually write my own html?? Oh noe…
As a European its always very hypocritical when the US warns of Huawei Hardware when there were cases of tampered Cisco Hardware. I would not really trust the Chinese Router either way but when it comes to AI models…the local chinese bias model seems to be the better choice than the one that lives on a server over there in the USA. As long as I dont ask for historical data, social studies or the Tianman Square it should be good no? What could it do? Hide the communist manifest (do they still care about that anyway?) in some bash script?
make sure you don’t ask the american LLM about Palestine 😹
You can also use Heretic to decensor open-weight models, and there aren’t many of those being produced by American labs.
Personally, I would prefer the Swiss’s open Apertus models to become competitive, but they are at least a year and a half behind the Chinese at this point IMO. It will be a while before European offerings are good enough.
i use HauHauCS’s models for that which are better then Heretic
Can’t be worse than Claude and Grok. Both of which the US seems to have no issue with.
For context, Grok called itself mecha hitler, while Claude is run by a doomsday cult that worships AI and wants it to destroy the world. Both are far worse than I could imagine any Chinese AI being.
Its probably less the performance of the AI model, more the data leaked to the Chinese government.
Its probably less the performance of the AI model, more the data leaked to the Chinese government.
Grok is owned by a Russian asset who stole America’s health and social security data. And it’s a garbage AI.
The DHS is less xenophobic about South Africans than Chinese.
the only cons with Deepseek is it trains on everything you tell it tbh
also pros Deepseek is open-weights which is better then Grok only older models and Claude does not make open-weights at all
They must also rip out the red leds from their screen, because that can combine to pink, like in the LGBTQ flag, which doesn’t align with government policy.
It’s self hosted tho?? I don’t get it. Isn’t remote-togglable stuff way more dangerous, like I don’t know, the Star Link?
You realize they want money to be spent, right?
This ain’t about saving money.
Wdym they have a online offering
I would assume the gov agencies self host their instances due to sensitive data. But even if they don’t, why not switch to self host?
Because idk, you need a good model and you dont have capex to buy a shitload of gpus
But yeah, ppl probably run small local models and they hate that too i can imagine. They want all the traffic to go to openai
Self hosting costs money in the form of computing hardware
nah they are using anthropoic, openai, MS, and google even.
Honestly, reasonable.
I mean, I sure don’t trust OpenAI or Anthropic any more than I trust DeepSeek, but that’s because I don’t have legal control over them. The US government does, and it is correct to ban tools that are controlled by adversaries.
You can run deepseek locally, have transparency into the tools and models.
There is less government control with deepseek than American born AI.
Everyone conflates “AI model” with “AI platform.”
The truth of the matter is that if you are interfacing with the DeepSeek service, you don’t have control of the version of the model being run, the hidden system prompts, or any co-agents that might modify output before its sent to the user/service.
The model itself could be completely innocent. It may not have had any weights modified and was trained on honest greenfield data. But we control literally nothing about the environment it is running inside of.
Meanwhile Elon Musk has a panic attack and reprograms Grok Everytime it tells the truth because it’s “Woke”.
Don’t act like US based models are not being manipulated as fuck. Hell it’s probably the ultimate end game for these pricks. Get these things mainstream for “all knowledge” and then “Hey ChatGPT, tell me about slavery?” “Ok user, slavery was a tool used by the benevolent superior white men to help grateful black people find work and shelter.” Except it doesn’t say “black people” in the actual future timeline.
I’m not acting like they’re innocent either, I don’t and won’t ever use any of them. But if you are thinking about it from the perspective of geopolitical warfare, the Chinese should doubly not trust grok and the US should doubly not trust DeepSeek. Using the adversaries tool is a fool’s mistake outside of R&D.
But the main reason for this move isn’t concerns about privacy or spying. That’s only the cover story. It’s to make sure taxpayer money keeps flowing to US companies like Anthropic, Musk, and OpenAI which are only marginally better in most tests than Chinese models and are many times the cost in some cases. Propping up the bubble a bit longer so that more taxpayer money can be stolen is the number one priority. Do you honestly think Trump gives a damn about the security of the US?
Two things can be true
Reading be hard
isnt he panicking over zelenksy bombing the russian datanceter right now.
There are US hosted deepseek llm access points. openrouter, zen are local, but its more groups that can spy/train from you.
you don’t have control of the version of the model being run, the hidden system prompts, or any co-agents that might modify output before its sent to the user/service.
A bit of a paranoid concern. Your worry is that service may try to lie to you, perhaps because of your nationality. There have been quality issues from US served Chinese models. Completely free access points will get ultra compressed/quantized/lobotomized in order to be cheapest to serve but also give a poor impression of the model. Q4 quantization is considered a good balance between cheaper and full quality. But that will still make more mistakes, and US services can fake providing full precision. Every model can choose to lie to you, and you can screen shot it. Using a secret liar model in place of advertised model is possible from any provider, but they are incentivized to be as good as they can be.
That doesn’t apply for open models that you can run locally, like Deepseek.
It’s still a risk. You can train an LLM to respond maliciously to a very narrow range of requests that would not be easy to spot in general testing. If you know how the LLM is used, you can target that very specifically. It doesn’t have to be as blatant as “always deny the Tiananmen Square massacre ever happened”.
Given how thoughtlessly people use AI, and how opaque some of the systems are, I think this is a legitimate concern.
And yes, every government outside America should ban the use of ChatGPT and Claude. Obviously.
But you don’t need to rely on testing. Deepseek is open source so they could just go through the code
Well, open weight is not open source. There’s a lot of black box still left in those models.
It’s still far more open to investigation and analysis than the likes of ChatGPT, though. Lots of work has been done to allow some degree of investigation into how the “brains” of an LLM work and in the case of Deepseek you can actually apply those techniques to the model you know you’re using.
there is no comparison between software and services, though. self-hosting deepseek might be, subjectively, the wrong choice for your lab, but chatgpt is wrong for everyone, every human and every animal and plant too
this is coming from a k2 7b user btw. it’s good enough for me!
edit: i replied to the wrong fucking comment and I’m too angry to fix it
Thanks for at least not being so angry you didn’t bother adding that edit, I was very “wut?” There. :)
I’m not aware of any realistic way to vet published model weights like that.
It works for some open-source software, because you can download the source and compile it yourself. But most of the “open” models don’t publish their entire training set, or enough details on the training for it to be truly reproducible. DeepSeek doesn’t publish their data sets, and they probably couldn’t, legally, since they “collaborate with third-party data providers to obtain proprietary datasets”. And even if they did, it would be prohibitively expensive to train it on your own, unlike simply compiling open-source code.
The actual code for training LLMs is relatively trivial. The secret sauce is the training data and fine-tuning.
You have agents classify what each portion of the network responds to and how. One compares it to filling out a very large crossword puzzle: https://already-clean.goodfire.com/
That’s the theory anways, not sure if it will stand the test of time…
Sounds interesting, thanks for the link. It’s a little over my head tbh but I’ll definitely take a closer look.
There is no such thing as transparency into large language models. It’s a billion numbers that don’t mean jack shit to anyone, human or machine. In other words: it’s Chinese to us.
Isn’t that the same argument they made for TikTok though? They are just trying to eliminate competition to prop up companies here.
That part tho
Yes and yes.
I don’t think TikTok is really comparable though, unless there are somehow people using it for important government work.
There are outreach campaigns telling you to get vaccines and warning you about recalls, or at least there used to be. Not really comparable, but still important government work that could benefit from TikTok.
Good thing that you can’t buy the reasonably priced Chinese EV’s in the US, like in the rest of the world, or the mobile phones, or the network gear. All banned by the US.
It is best you buy only American and see how far you get, especially in innovation.
Why do you think the US government is removing all options for competition in any field?
It’s incredibly ironic that China used to have really really bad EVs so they let other brands (mostly Tesla at the time) enter their formerly closed markets to boost competition, low and behold, it worked so well the US has now blocked them because the US is afraid of competition.
i bought OPR last year as my phone because its battery reliability and affordable
Exactly the reason why every country outside of America should drop all support for their IP and military purchases.
It’s not safe when dealing with a facist regime that controls the on/off switch to their own military operations and defences.
We’ve seen with Ukraine and the U.S. throttling access to starlink, and removing their sales or munitions to Ukraine to defend themselves again another rogue state.
Rogue states cannot be trusted.
musk told trump to throttle Starshield, which is just rebranded from starlink probably he gave info on how to mess with it without directly interfering.
I have no idea why anyone would be using the online version of deepseek or qwen or glm or kimi in the us gov, surely they would be using chatgpt or claude?
i was so confused im like who cares if they’re running deepseek in lm studio or on their own servers but then they mentioned the sign up process
It says DeepSeek’s sign-up process collects personal information and may “track and digitally fingerprint users”, and that its apps work by having people type in information, which could mean sensitive data from US government staff or contractors ends up in China.
im like right, the online version, that makes sense
I know people in state gov who couldn’t even use CoPilot (blocked by firewall) until an explicit contract was put in place with Microsoft about data handling. (On top of their existing Microsoft contracts.) All other LLM’s, ML, or “AI” are blocked unless approved, as well.
Seems like the federal government does not have its shit together.
I worked for a Microsoft studio for a while that took forever to allow Copilot to be used internally, and they were all Microsoft. It was very frustrating, especially given that permission was finally granted for Copilot use just a matter of weeks before the whole studio got abruptly shut down anyway.
Corporate bureaucracy. Weird and dumb.
Military should not depend on the imported (non-domestic) stuff. Supply chains must be 100% domeatic for the military.
Those are military requiremets.
This move is an overreach into the civilian/private sphere, so nakedly designed not for security, but for privileging the already morbidly overpriviledged domestic billionaires.
Also, with the rabid billionaires in control anyone can become an adversary overnight. Trump, with the presidential cabinet filled to the brim with billionaires and bootlickers, made Mexico, Canada, Denmark, and even the whole of the EU into adversaries, when they weren’t before. The word “adversary” is meaningless these days.
I trust deepseek because I can run it fully offline on my own machine
That must mean it’s useful.
Super Intelligence good, China Intelligence bad!
Happy Cake Day 🍰
